China's AI Breakthrough, The Hidden AI Debt, AI Bubble Warning, and More

Okay, so John, um, maybe you can take, take us through, uh, what happened with the, uh, Kimi K2, Kimi K3 Chinese open, open weight model that, that landed this week.

Yeah, absolutely.

Uh, boy, and listening to that list of things you were covering today, we've, we've got some breadth here.

We got some work to do, Greg.

So, um, so, uh, he- here's the, the, the summary on the open weight models and the Chinese, uh, models that have come out.

I think somewhat unsurprisingly, the, uh, the models that are, that are coming out a- at each iteration or generation are surprising everybody in terms of how good they are and how much better they, uh, they, they are than expected.

We've seen this just time and time again, where every time a new model comes out, uh, either from the US company or the Chinese side, there's all of this kind of mind-blown, uh, responsiveness.

I think at some point we have to just, uh, uh, realize that this is, this is, uh, incredible and, and, and, uh, going to continue to be so.

DeepSeek and Manus were that way for people, you know, when they came out.

Yeah, they really were.

They really were.

So, so at some level I would say this is sort of more of the same.

Everybody's mind is blown all over again.

The implications are, are, are pretty interesting, and what may be going on behind the scenes, also pretty interesting.

So, so the, the conjecture here is, is a lot of these companies are looking to figure out how to continue to do their AI coding and development, but without the huge bills that they're getting, in particular from Claude.

Um, do these models offer an alternative at, you know, 10%, 5%, or even less the cost of using, uh, Claude's APIs? And this is a, a, a, a, a general, um, uh, you know, a sort of a concern for the US AI companies who have had to have these pretty aggressive API prices in order to start getting at least some revenue.

So, you know, what's going on here? Well, it could be, uh, potentially very disruptive.

Here's a nuance that I think gets lost that will have to be solved, but when you're talking about models that are 5% as ex- as expensive, there's a lot of incentive to solve them.

Uh, you can't just take something like Claude Code and take out their models and plug in, uh, you know, Kimi or, or, or one of the other Chinese, uh, open, open weight models.

It's not the way that it works.

Um, Claude wouldn't want you to do that.

And in fact, there's a lot more than just the model that is under the hood that makes these things go.

On the coding side, it's called a harness, and a harness is really the set of instructions that tells the model how to approach coding projects.

And I will say Claude Code and their harness is exceptionally good, and I would not myself want to be in a position where I was trying to recreate that harness and plug a Chinese model into it.

I, I, you know, anecdotally happen to think that, you know, more than half of the capabilities that something like Claude Code has have to do with the harness as opposed to the underlying model, and that may become even more important as these models achieve parity.